5 ms·
They're discrete with such high cardinality that successful approaches will likely model them assuming they're basically continuous. Neural network layer activa
by gradys 9y ago
They're discrete with such high cardinality that successful approaches will likely model them assuming they're basically continuous. Neural network layer activations are also discrete after all, but they're often 256+ dimensional vectors of float32s or float16s.
- rjeli 9y agoWell, WaveNet[0] outputs audio in the time (not freq.) domain using PixelCNN, so it's not unthinkable. https://deepmind.com/blog/wavenet-generative-model-raw-audio/ https://deepmind.com/blog/wavenet-generative-model-raw-audio...